Good morning.
I asked Dall-E to create a figure, and show it from top to bottom. It gave me two images, each of which was incomplete. Image 1 cut off around the knees, and Image 2 cut off below the belt.
When I told Dall-E about this, it insisted that I was mistaken, and that it had created the full images.
When I told Dall-E that even after downloading the images, they were exactly as I’d described them, I asked Dall-E on what to tell someone about this.
Dall-E gave me the following…
[ * Suggested Improvements:
- Enhanced Image Analysis: Suggest improvements in the system’s ability to analyze and understand the full scope of image generation requests, ensuring that full-figure images truly depict the character from head to toe.
- Clarification Features: Recommend the implementation of a feature that allows users to specify or confirm critical aspects of their request (e.g., ensuring the full figure is visible) before the image generation process begins.
- User Feedback Mechanism for Generated Images: Propose a feedback mechanism that allows users to report specific issues with generated images directly, aiding the system in learning from errors and improving over time.]
For the record, I enjoy Dall-E. I’m not angry. I just thought if this is the advice that Dall-E itself is providing, the people who refine the program might like to know about it.
Also, I’m not expecting any replies on this. Just passing it along.
Best to you and your loved ones.
You describe ChatGPT. It can only send image descriptions, it can’t see the images or figure out why the images are made the way they are.
You can use prompting techniques (“depicted head to toe in a full-body-length image taken from 8 meters away”), and also ask for the tall aspect ratio image. However 50% of the time you can get the desired content – but not the desired rotation of the base image size by the internal techniques of the image creator…
Good afternoon, _i.
I don’t mind if the image is rotated, since I can fix that after downloading it. It’s that I gave Dall-E an instruction to show the figure from head-to-toe, only got a partial image, and then had the following exchange…
[You
Please take both images, and show them from the top of the character to the bottom.
DALL-E
The images already depict the characters from top to bottom, showing the entire figure of the female heroine as per your request. You can view the full length of each character in the provided images. If you’re looking for a different angle or perspective, please let me know so I can assist you further.]
Dall-E couldn’t take another look at them to see that it was wrong, but somehow “knew” that it was right.
In fact, earlier on, I asked Dall-E to take my entire prompt and rewrite it so to increase the likelihood that Dall-E would give me what I wanted (I was getting crosshatching that I didn’t ask for). So Dall-E was failing to respond accurately to a prompt that it itself had created for me.
However, since you brought it up…and any advice is welcome (I’m new at using this program)…what did you mean by “ask for the tall aspect ratio image?”
Best to you and your loved ones.
Just give certain keyword sequences to ensure the image is depicted as you want.
Here’s an example starting a “new chat” with no custom instructions.
Thanks for the reply. I’ll experiment with something like your prompt later on.